Indispensable AI
AI’s five-layer architecture, and why the Application Layer matters most.
The AI Stack
Artificial intelligence is a stack: energy, silicon, cloud, models, and applications. Each has its own economics, competitive dynamics, and challenges. Capital, regulation, and strategic attention concentrate at different layers at different times. Mistaking one layer for the whole industry causes confusion, misrepresentation, bad decisions, and misguided capital allocations.
- Energy
Power generation, transmission, and grid capacity are critical constraints under extreme pressure to develop capacity and capability. This will be longer and more inefficient than the most optimistic projections. Regulatory requirements and capital investment are not as freely available as assumed.
The data center buildout has converted electrons into a critical input for AI. Hyperscalers signing nuclear power purchase agreements, sovereign capital funding fusion startups, and grid operators rationing interconnection queues mean one thing:
AI is a physical, electricity-driven industry.
Electricity generation and transmission will not come easily. Permitting new energy-generating facilities and their transmission is hopelessly bogged down in local, state, and federal regulations. Current energy customers are only beginning to protest limited supply and rising prices, as supply is being refocused to serve hyperscalers.
Nuclear power takes many years to develop and is not simply “turning on a switch” to provide more energy. Fusion is a wonderful dream that is unlikely to manifest anytime soon, especially within a reasonable timeframe (or in my lifetime).
Energy is the foundation of AI, and a critical constraint. Power, efficiency, and AI-driven solutions will help solve this issue. It is a typically overlooked component of AI processing. Solutions are coming, and energy is available for the AI engine to continue. Luckily, more efficiency is on the way.
- Silicon
This layer consists of some of the most innovative and disruptive companies in industrial history.
Logic chips, memory, packaging, and the tooling that produces them constitute a critical industry sector in which several irreplaceable companies with dominant positions hold long-term competitive advantage. NVIDIA, TSMC, ASML, SK Hynix, Samsung, AMD, Broadcom, and a tightening list of national champions sit at this layer.
The competitive structure is brutal, capital-intensive, and geopolitically encumbered. It is also the layer where the United States retains the deepest lead and the most defensible chokepoints.
The industry is evolving with new companies like Cerebus and Groq, and reemerging competitors like Intel, all focusing on chip designs to make inference more efficient. It will be challenging to displace ecosystems built by Nvidia and others, but demand is so great that the industry has created enormous capacity for new entrants and previous laggards.
The constraint is production, resting in one single chokepoint in Taiwan. TSMC drives the entire industry, is scaling with more than $20 billion annually in capital expenditure, and still cannot keep up with demand.
Trillions of dollars of value are created at this layer by some of the world’s savviest and fiercely competitive companies. This layer is innovative, robust, and driving the AI industry through processing capabilities and tools essential for product design and service delivery.
- The Cloud
The cloud is becoming a laundry list of formidable competitors spending unimaginable sums to build out AI processing facilities. Amazon Web Services, Microsoft Azure, and Google Cloud Platform, joined by Oracle, CoreWeave, and a host of others, translate silicon into globally distributed compute.
This is the layer that transmits AI workloads. This is the global infrastructure of artificial intelligence workloads. The hyperscalers collect an infrastructure tax on every token generated above them, and they spend hundreds of billions of dollars annually to maintain that position.
The incumbents’ competitive advantage and sustainability are insurmountable. The intense processing and capital expenditures create one of the largest moats in industrial history. It’s an arms race to continue providing services, but only the largest armies can sustain themselves.
Here is where vertical integration really matters. Google is vertically integrated in chips, cloud services, and, as we will see, in models and increasingly helpful in applications. Amazon is dominant in the cloud here, but it is also developing its own silicon. Microsoft has cloud positioning and is trying to continue owning the application layer (which is increasingly vulnerable). It is trying to gain access to models via its investment in OpenAI. This has made Microsoft vulnerable, and it is looking for other alternatives, including Anthropic.
The wisdom of these choices remains to be seen, but these are the players, and their capital spending drives value for chip manufacturers and enables the model’s capabilities.
- The Models
OpenAI, Anthropic, Google DeepMind, Meta, and xAI, together with the open-weight challengers led by DeepSeek, Qwen, Mistral, and Llama derivatives, produce the raw cognitive capability that the rest of the stack depends on.
Models are the engine.
The models have received disproportionate attention, and while their functionality is essential to building economic value for applications and to driving AI deployment, the arms race among model manufacturers is creating increasingly challenging competitive distinctions.
More and more, they are becoming interoperable, and the standalone models from OpenAI, Anthropic, Perplexity, and others are facing challenges of enormous capital expenditure, an inability to be distinctive with a unique value proposition, and dependence on cloud services, semiconductors, and distribution from other vertically integrated companies.
The more powerful competitors that are vertically integrated, Google, Amazon, and Microsoft, have a distinct competitive advantage for sustainability here. It is somewhat unlikely that standalone model companies are truly sustainable. Given the capital expenditures required to generate revenue from these models, it is expected that consolidation and combinations with vertically integrated companies will occur. Models are not going away and are extremely valuable, but as the world moves from training to inference, the distinction between them is increasingly blurred.
- The Application Layer
This is the sustainable value-add layer where intelligence is converted into decisions, workflows, products, and outcomes that customers actually want and will pay for.
It is still emerging; a good analogy is that the application layer is in the second inning of a baseball game. But this is the layer where the largest share of long-term economic value will be captured.
The engine that drives value for the entire AI stack is in the application layer. All the other layers form a foundation for producing a product, but now the critical question is: Is there any value in what is being produced?
None of the other layers matter if the application layer doesn’t deliver. Energy, silicon, cloud, and models are means for applications.
Extraordinary infrastructure is being built, leading-edge technology is driving processing, and together they create a platform for valuable applications.
The infrastructure and tools present an enormous opportunity at the application layer to create extraordinary value. It’s unpredictable, and the winners are still far from clear, but this is AI’s future.
Applications Win
We’ve Seen this Movie
It’s happened before. The infrastructure builders enable the platform; the application builders capture the value. Railroads enabled industrialization, but the largest fortunes accrued to manufacturers and distributors who used the rails. Telecommunications networks enabled the internet, but the platforms and applications built on top dominated value creation.
Cloud infrastructure produced trillion-dollar businesses for AWS, Azure, and Google Cloud, but the valuable third-party companies of the cloud era — Salesforce, Workday, ServiceNow, Datadog, Shopify — were applications that sat above the infrastructure and captured premium pricing by focusing on specific applications that dominated their sectors.
It will be no different for AI. Real opportunities will focus on value-added products and services delivered on top of infrastructure. The opportunities are enormous, and the total addressable market is essentially the global economy.
A Token Is Not a Product
Models produce tokens. Tokens are an intermediate good. Customers do not buy tokens; they buy completed legal work, approved loans, validated drug candidates, optimized energy grids, scheduled cargo, diagnosed patients, and shipped code. The opportunity is the spread between the cost of producing tokens and the value of those products.
Model makers capture some of it through their first-party products. Cloud providers capture a layer through hosting fees. But the deepest, most defensible value is provided to specific industries. Specific applications require qualified data sets, industry expertise, regulatory compliance, workflow integration, and the creation of sustainable value for the customer.
It’s Not a Chatbot
A generic chatbot wrapped around an LLM is not defensible. An example from healthcare can illustrate this point. This industry is highly regulated and requires deep institutional knowledge and experience. A generic chat by an “AI agent” cannot serve this industry in a sustainable and valuable way.
A company woven into a regulated clinical workflow, with years of institutional context, validated audit trails, and integration into the systems of record on which a health system actually runs, is a different kind of asset. This market opportunity is still developing and cannot be cracked by an all-purpose agent from a Mac mini.
The application layer is still being defined because incumbent software companies, such as Salesforce, Atlassian, and SAP, still offer valuable enterprise-specific software. But they are under threat from agentic AI. But institutional knowledge, customer interface, and comprehensive service are not easily delivered by an AI agent.
Work may not be performed inside dedicated software (for example, Claude can now essentially create what would otherwise be found in Word or Excel). The source code is facing diminishing value because AI models can now produce the same graphics and functionality. However, software still needs to be comprehensive, interactive, and responsive to customer needs. Agents do not do this easily; AI and the application layer still offer significant opportunities for software incumbents. But it will be a dynamic, competitive environment from now on.
Industries, Applications, and Opportunities
- Life Sciences, Healthcare, and Biotechnology
Biotechnology is an industry of failure, wasted capital, extended time frames, frustration, and a profoundly inefficient ability to serve unmet medical needs.
New therapies now cost $2 billion and counting, and take decades to deliver. Diagnostic error contributes to an estimated 800,000 cases of death or permanent disability annually in the United States alone. Administrative overhead consumes roughly a quarter of US healthcare spending. This is not sustainable if society is going to have useful, impactful health care without bankrupting itself. Each of these is now an AI-application target, producing defensible, innovative performance.
Drug discovery
Google’s AlphaFold gets significant headlines and is the inspiration for generating potential drug candidates from an otherwise unimaginable data set. Turning these candidates into approved therapies is still a big jump, but the shift is on.
Several companies are generating clinical-stage drugs discovered using AI-native tools and applications. By the start of 2026, AI-native biotechs had multiple AI-designed candidates in human trials, with the expectation of having three or more in the clinic during the year.
The therapeutic targets are previously intractable challenges, including ALS, autoimmune disease, and oncology. The large pharmaceutical companies understand this transition and are dedicating billions of dollars to partner with AI companies developing algorithmic protein engineering, AI-based drug discovery, AI-based therapeutic testing, and other potential efficiencies to bring effective drugs to market sooner.
Diagnostics and clinical decision support
Artificial intelligence is replacing medical practitioners and experts. Still, these medical experts are using their experience and awareness in partnership with AI applications to be more effective and enhance the quality of patient interaction and care.
An example is AI applied to medical imaging. This will not replace radiologists but make radiologists more effective. Fundamentally, this sector is the blueprint for AI applications throughout all industries. It doesn’t replace so much as enhance the capabilities of all professionals within that sector. This is magnified in medical care, where patient care, time for thoughtful consideration, and case review are greatly enhanced by AI, which processes data and provides results. It is still experiencing medical awareness among the practitioners who make decisions. But this is where AI applications and proven industry, its practitioners, and its patients come together.
These applications improve outcomes while still driving down costs. This is an unprecedented solution for health services.
Healthcare operations
The seemingly intractable administrative and bureaucratic juggernaut of healthcare operations is likely to be the first target.
Prior authorization, claims adjudication, revenue-cycle management, scheduling, and the dense paperwork that consumes a quarter of US healthcare spending are precisely the target of AI applications.
Digital-health venture funding reached $4 billion in the first quarter of 2026, the strongest opening quarter since the pandemic peak, with average deal size at its highest level since late 2021. This capital flow is likely only the beginning of a focus on a sector where AI applications have demonstrable and large-scale economic impact.
2. Financial Services
Finance and financial services are typically among the first to adopt technological innovation. The combination of enormous data, intense processing, and value creation from knowledge, coordination, and effective delivery of products and services makes the financial industry one of the first and most extensive adopters of artificial intelligence.
Finance has always been a quantitative industry layered on top of unstructured data, including financial statements, credit memos, regulatory filings, research reports, trader chats, and contract clauses. AI enables the quantitative management of unstructured data. The result is an unprecedented organization of information into manageable, thoughtful structures, which is driving the most aggressive enterprise adoption curve of any sector.
By the end of 2025, more than 90% of major financial services firms were piloting or deploying generative AI in core business functions. Roughly half of asset managers either use AI or expect to do so within twelve months.
Over eighty percent of NYSE trades are now executed by algorithms. AI is permeating and integrating all aspects of finance and financial markets. It is impacting pricing, risk management, execution, regulation, and financial innovation.
Trading, risk, and asset management
AI-driven systems now perform at scales human analysts cannot match, integrating data sources into effective models at a scale that required dedicated research teams. Now, it requires a well-designed AI application with oversight.
Risk management has followed. Sixty-eight percent of financial services firms now rank AI-driven risk, fraud detection, credit analysis, and compliance as top strategic priorities.
Business models are being reinvented with an AI focus. AI-powered applications are being deployed in every sector of finance. This is the vanguard of value-added applications to be developed further from here.
Banking and consumer finance
Personalized advisory, once the preserve of high-net-worth clients, is now available for all financial services customers. Ranging from lending, business services, investment management, credit services, underwriting, or transactional data analysis, services are now more broadly available and cost-effective for all bank customers. This is fundamentally disrupting the services and banking industry.
AI-driven underwriting, for example, is compressing the time and data required for underwriting and access to capital. This is a fundamental driver of economic development and entrepreneurship.
The impact is global. Now it is cost-effective to serve customers in emerging markets, catalyzing economic growth and business development.
Credit drives the world. AI-driven applications will be more valuable and provide greater assistance to economies, business owners, executives, and entrepreneurs globally than any other program could.
Regulation
The industry is ahead of regulation once again. Any friction or delay caused by regulatory compliance slows the economy. Finance is the intermediary of industry. The AI application layer can provide more effective products and services, but regulation can grind this to a halt.
Roughly two hundred state-level AI bills are being tracked in the United States in 2026 alone. Regulators know something should be done, but, obviously, they are not really sure what that should be.
We know cybersecurity, consumer protection, balancing automation with human oversight, and other issues are complex and unresolved. The industry will move faster than regulation can keep up.
Failure in finance can be catastrophic. That still doesn’t mean we will have clarity and efficiency anytime soon.
3. Energy
AI is the largest new source of electricity demand in a generation, and the most powerful new tool for managing the grid that must serve it.
Data-center power consumption is forecast to reach roughly six percent of US electricity by the end of 2026 and to approach ten percent by the end of the decade. The International Energy Agency projects a 40% surge in global electricity demand over the next decade.
Without electrons, the chips do not run.
Grids and Demand
Grid optimization and demand response are two of the most underappreciated yet essential developments in AI applications.
AI applications are now embedded in the operations of essentially every modern grid operator. The electric grid is no longer a static network. It is becoming a software design system that manages energy, grid deployment, transmission, and balancing. Already, systems using AI have demonstrated 20% reduction in electricity costs.
As society becomes increasingly electrified via mobility, data centers, and other applications, AI will be essential for conservation, design, delivery, and innovation.
Energy is the foundation for the global economy. As a Nobel prize-winning economist has said, “power is prosperity.” Energy management and AI applications for energy drive global prosperity.
Generation: from renewables to fusion (perhaps)
Leading global energy demand will require innovation in materials, systems, processing, transmission, and generation. AI can simulate different materials for solar panels, discover new battery chemistries, new materials for more efficient transmission, and design small modular nuclear reactors. All of this can lead to energy efficiency, carbon-efficient generation, and effective transmission.
Nuclear fusion, a dream that has been 20 years away for the last 50 years, may enable endless clean energy. While this still seems scientifically challenging beyond human capability, if it’s going to happen at all, it’ll be because AI simulated the intense combination of scientific breakthroughs and plasma simulations required. Venture capital is pouring money into this, and while its outcome is uncertain, capital, intelligence, and AI applications will be the condition precedent if it’s going to happen at all.
The renewed interest in energy is driving innovation, venture capital, scientific study, and interdisciplinary cooperation on an unprecedented scale. This innovation is essential for efficient, plentiful energy to drive the infrastructure of the new economy. Artificial intelligence won’t happen without plentiful energy at its foundation.
Trillions of dollars are at stake, from semiconductor manufacturing to cloud services infrastructure and data processing, and it all depends on plentiful energy.
Expect AI-driven innovation to be unprecedented.
- Mobility, Robotics, and Advanced Manufacturing
AI in the physical world is developing exponentially. However, exciting developments in the lab are not quite the same as commercial deployment. Still, so much attention is being paid to the efficiency and capability that AI is bringing that there is much excitement and hyperbole around what’s possible.
This is a big difference between a humanoid robot dancing on the stage and working in an intense factory setting with little margin for error. However, mobility and autonomous industrial processes are receiving enormous capital investments and global industrial attention. Applications, value-added products, and services will expand from here.
The application of AI to the physical world will be dramatic.
Autonomous Mobility
Autonomous mobility is here. Improved safety, with up to 90% reduction in vehicular accidents, is demonstrable and desirable. Access to mobility for the aging population, efficiency in trucking and other delivery services, and global scale for mobility-as-a-service are emerging. AI-driven applications with dense, sector-specific datasets will enable more efficient and pervasive mobility services.
General autonomy in the products and services related to that will be AI-driven, but the path is still long, and operational design and commercial viability are still developing. Competition among Waymo, Tesla, Uber, and others will drive down costs and increase efficiency, but the full deployment of true value-added services remains in its infancy. But the sector will be impactful.
AI will drive mobility-as-a-service and change driving and transportation forever.
Industrial and Humanoid Robotics
Robotics is given a lot of attention. Some of this is hype, from dancing robots to aerial drone shows. But underneath, there is a disruption in industrial systems and applications. It is not simply applying AI software to hardware and flipping a switch. Systems engineering is challenging and requires robust “six nines” reliability, but the work and capital required for research, innovation, and development are underway.
The global market for industrial robot installations reached an all-time high of over $20 billion. Training robots and deploying them in the physical world without catastrophic failure or inefficient performance still requires more time and development, regardless of what YouTube videos may say. Deployments are continuing, testing is ongoing, and robust applications are increasing. While none of this is around the corner, the process and the direction are irreversible.
Humanoid robotics is a more curious development. Industrial settings do not require systems, processes, or equipment to resemble humans, yet over $5 billion in venture capital was raised last year for humanoid-robot development. As an example, it’s another hyperbolic announcement from Tesla. Whether we all need an Optimus Prime in our homes is another question.
The thesis that they can operate in an unstructured environment is intriguing because structured environments require different kinds of capital. Whether there is a consumer deployment or a need for a humanoid form factor, or why a robot needs to be active in an unstructured environment, is yet to be determined.
The fundamental point is that the AI application layer, when applied to thoughtful hardware design and rigorously tested, can produce unique applications. The commercial value is something the market can determine. Right now, the jury is out, regardless of all the breathless announcements.
Advanced Manufacturing and the Agentic Factory
Innovative design, digital twins, vertical AI applications, and other advancements are possible as advanced AI-driven insights drive efficiency and innovative systems engineering.
Companies are integrating physical AI into manufacturing through digital-twin platforms that simulate and validate entire factory operations before physical implementation. AI welding, AI finishing, AI assembly, and AI inspection are arriving as pre-trained, pre-integrated modules that deliver immediate productivity.
Industrial automation will grow as AI delivers demonstrable manufacturing benefits in throughput, quality, and cycle time. The combination of a structured environment and data intensity is AI’s strength and will impact systems engineering, design, and manufacturing.
5. Aerospace and Defense
Artificial intelligence is impacting global defense. The cost-effectiveness of new weapons, software-guided drones, and other weaponry is disrupting budgets. This is best demonstrated by the anachronistic strategy of using a $2 million missile to shoot down a $40,000 drone.
When the US Defense Department budget (which may be as much as $1.5 trillion) realizes that much of it may be spent on ineffective weaponry, a substantial portion of the budget is now being allocated to AI.
This is a global phenomenon. The tragedies of the war in Ukraine and Iran, as well as other never-ending global conflicts in Africa and the Middle East, show that cost-effectiveness has now enabled more combatants and therefore requires countermeasures that will be fundamentally based on AI applications.
The implications are obvious because the pace at which AI-driven weaponry can be deployed, maintained, and replaced is meteoric compared with traditional procurement systems. On top of this, there are autonomous mission planning, real-time collision avoidance systems, geospatial targeting, predictive maintenance, and AI-assisted certifications of complex systems, which are converting traditional aerospace and defense from a hardware-dominated industry into a software-and-data industry whose hardware is increasingly software-defined.
Commercial aviation is following. AI is being deployed for crew scheduling, fleet optimization, predictive maintenance, and passenger experience. The efficiency gains from AI services in complex scheduling for equipment and personnel can be dramatic. Perhaps more interestingly, advanced air mobility, including autonomous cargo drones, air carriers, localized passenger travel, and other AI-driven mobility air services, can generate significant revenue opportunities. Additional capital investment is driving this further.
6. Communications, Information, and Cybersecurity
One of the most consequential applications of AI is the production, distribution, and verification of information.
The AI application layer is restructuring search, customer support, and the entire “knowledge economy.”
Cybersecurity is a problem and a solution. Recent headlines about Anthropic’s Mythos, which can either crack every bit of software security or solve all software vulnerabilities, are still up for debate. The clear issue is that AI applications threaten software security. It can be adversarial or defensive. Regardless, it will be critical.
Software ate the world, and now the world needs software to keep the software from eating it all over again.
But this time, software is the adversary of software. What it means is the world runs on software, and that software is now extremely vulnerable.
Essentially, there is no longer a security perimeter around the network edge. Security is a substantially harder problem, and cybersecurity will be a critical issue for all AI applications.
7. Applications
AI applications thrive in an environment with defined, high-quality data that is plentiful and accessible. But, controlled and secure, preventing hallucinations, gross errors, or other misleading results. AI can create powerful applications when combined with organizational knowledge, history, a clear business case with economic rationality, and effectiveness in highly competitive environments.
Healthcare, financial services, defense, and energy all share these characteristics. So do segments of legal, accounting, and compliance work. Where these conditions are present, the AI application case for highly competitive industries is self-evident.
It’s Not for Everyone – Yet
However, this does not apply to all industries and all applications. Many times, there is inefficiency, confusion, and no real value case. This may change as AI tools develop, but AI is not a panacea. Where it is impactful, it is significantly impactful, and those deployments will accelerate.
Fortunately for the AI application layer, some of the world’s most important industries have characteristics that make AI applications highly valuable. These industries will be disrupted, dramatic changes in workflows and personnel will occur, and new creative opportunities will emerge unpredictably but inexorably.
The Innovator’s Dilemma
The winners build AI-native architectures rather than retrofitting legacy systems. Incumbent software companies attempting to add AI to existing seat-based products face the dilemma that defines every disruption: protect the existing revenue model or cannibalize it before someone else does. The companies that hesitate become case studies. Those that accept cannibalization survive.
The new entrants built from the first line of code on the assumption that an agent, not an operator, will execute the workflow, capturing most of the value that the incumbents lose.
The human element cannot be underestimated. Many AI application trials fail to deliver measurable results. It’s not necessarily a verdict on the technology so much as on the organization itself. People are hesitant to use AI; there’s a lot of confusion around it, and sometimes the applications themselves are inadequate. This is an inefficient process, with a longer timeframe than markets tend to expect, but adaptation and more effective solutions will come.
Shifting Balances
The center of gravity is shifting among all five layers of the AI stack. Energy, silicon, cloud, and models all represent unprecedented value creation and capital investment. There continue to be extraordinary capital allocation plans and dramatic infrastructure investment. All of this ultimately benefits the AI application layer. That center of gravity is moving towards the top of the AI stack.
Compute is not as scarce, although it is still scaling; it is not the critical bottleneck it was. Model differentiation is compressing. The model builders are converging and becoming less distinctive. Each announcement morphs OpenAI, Anthropic, and others’ offerings into something more indistinguishable. The model maker’s competitive advantage may be diminishing quickly, whereas the vertical players, Google, Microsoft, and Amazon, may have a more sustainable position.
Is Any of This Valuable?
But the real center of gravity is shifting to applications. Essentially, the question is now, what value does all this deliver? Energy, silicon, cloud, and models only serve to deliver that product. If it is not valuable, the entire stack is in trouble.
However, there is substantial evidence and a robust argument that we are at the beginning of an unprecedented value-creation curve. Built on the infrastructure and services provided by the other layers of the stack, the AI application layer will be globally transformative and disruptive. The direction is clear, the timing is uncertain. It will take much longer, but the impact will resonate for generations.
The constraints are imagination, execution, and the willingness to rebuild how work is done. That is also the opportunity. The application layer is where the next decade of value will be created, and where the architecture of the transformed economy will be transformed.
